New features based on the Cohen's class of bilinear time-frequency representations for speech recognition

Although short-time Fourier Analysis based features such as LPCC and MFCC have been widely used in state-of-the-art speech recognizers, the short-time analysis technique suffers from the well-known trade-off between time and frequency resolution and works under the assumption that speech signal is short-time stationary. This paper investigates an approach to use Cohen's class bilinear time-frequency distributions representing speech signal for speech recognition. Preliminary experiments show that the new feature can better represent speech signal and can improve the accuracy of a speech recognizer.